* Van Ly via Emacs news and miscellaneous discussions outside the scope of other Emacs mailing lists <emacs-tangents@gnu.org> [2025-03-05 04:12]: > > Jean Louis <b...@static.rcdrun.com> writes: > > >> > I wish to find out anything truly innovative that was invented by some > >> > "AI" Large Language Model (LLM). > >> > >> I have heard of a project that intends to analyze 130,000 documents in a > >> field, and if they are lucky, parts scattered in the documents may fit > >> together to give a new original insight in the data that was previously > >> overlooked. > > > > As you said, the large language model would learn from one hundred thirty > > thousand documents, and that means that it cannot learn something new to > > innovate something new. That means the original insight is not that much > > original, not so. > > I never said, the large language model would "learn", I get that LLM model a > form of jpeg zip compression of a cut off date internet state for information > retrieval in the most common use case. "AI" techniques have been recognized > for new original insight in two cases since AlphaGo.
Okay. When Large Language Model (LLM) analyzes 130,000 documents it means it is getting the information, thus uses that information. If you call it learning or absorbing, whatever, it took the information, it did not invent new information. You mentioned "new original insight in the data" -- that is matter of viewpoint and consciousness of the observer. If observer is not aware of his own information, of course, the output from model will give to the observer or operator, truly new and original insight that he overlooked. Every day I am using different memory for the LLM, and I include it as context for reason that I need specific results. For example, LLM doesn't know the date, but I can give to model on each request, either in a prompt or in a system message, accurate time and date. Imagine, it starts suddenly giving pretty accurate time, and it can calculate ages of people, and understand birthdays. For some business processes I am giving the stages or steps of the process, which differ from what model otherwise know and so I get always right answers that way. Through exercises of that type I can to a degree better understand how it works. The exercises gives me quite a feeling that model "knows" only that what it was given to know. In reality it knows nothing, we are only getting representations of vectors based on same data feed. Truly computer is there for reason to store information which we would have difficulties processing ourselves, and that information is mostly of less value for human life, as that what mostly matters we do everyday with or without computer. Original insights are basically "aha" moments of human who overlooked their own information. Large Language Model (LLM) helped human to get into it. My program for augmenting knowledge is called Hyperscope, Dynamic Knowledge Repository, as by Doug Engelbart. But we could use notion both of the telescope and microscope to understand some meanings. With the microscope, we are going to see let's say bacteria in blood and recognize what is going on. By purely seeing it optically, we do not invent anything and we cannot say that microscope is "originator" of some invention. Human had to observe it and make a conclusion of what was found. Did microscope invent the found bacteria? Telescope? It could see help human to see intereting things in space, and some people can be called "discoverer" of new planets, stars, life, whatever. But did the telescope discover it or human getting realization based on the tool? Microscope, telescope, hyperscope, those are tools, Doug Engelbart envisioned it long ago: https://dougengelbart.org/content/view/355/ They invent nothing new. With whatever alleged inventions, those tools do nothing and can't do nothing. There must be always human to make some assumptions, conclusions, after observations. Without observation there is no invention. I have thousands of documents, of course I could go through them and realize something "new" that I previously did not know, maybe that I am using specific words incorrectly. And I could have Large Language Model (LLM) parse documents and realize same thing. It is far from any thinking, it is computation based on given goals. By all means it is similar to computer program just with the NLP. The surprise that there is some invention by AI must be based on lack of consciousness of the observer on what is actually going on. Setting up rules and then asking model to computer by those rules will not give new inovation. Inovations are beyond what is known. But deception by human observe can be so strong. Best way to understand the text generator is to run it frequently. I see Large Language Model (LLM) as Collective IQ as envisioned by Doug Engelbart, see https://dougengelbart.org/content/view/247/ While they are not directly comparable, LLMs can be seen as tools that might contribute to raising a group's Collective IQ by providing enhanced capabilities for processing information, making connections, and facilitating communication. I was myself truly surprised that with just few words the model could basically predict what I want, and I have got a feeling "Oh, it can read my mind". With much usage over last 2 years, from smallest models on 4 GB GPU to largest online running models, I can see it is just matter of space (memory) that model needs to be more precise and more accurate, truly appearing highly assistive. And I wish I could find somewhere truly inovative, but I don't. -- Jean Louis --- via emacs-tangents mailing list (https://lists.gnu.org/mailman/listinfo/emacs-tangents)
